[Llama 3.1] Updates MLLOG tags#790
Conversation
|
MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅ |
|
@Elnifio can you verify that the logs from this reference indeed pass the latest checkers? |
|
|
||
| def log_validation_loss(self, metrics, step): | ||
| consumed_tokens = (step - self.init_global_step) * self.gbs * self.seq_len | ||
| consumed_tokens = step * self.gbs |
There was a problem hiding this comment.
shouldn't that be called consumed_samples?
There was a problem hiding this comment.
Updated in the latest commit. Thanks for catching that!
| def on_validation_start(self, trainer, pl_module): | ||
| mllogger.end(key=constants.BLOCK_STOP, metadata={'epoch_num': self.consumed_tokens(trainer)}) | ||
| mllogger.start(key=constants.EVAL_START, metadata={'epoch_num': self.consumed_tokens(trainer)}) | ||
| mllogger.end(key=constants.BLOCK_STOP, metadata={constants.SAMPLES_COUNT: self.consumed_tokens(trainer)}) |
There was a problem hiding this comment.
not sure where self.consumed_tokens() is defined, but does it return tokens or samples? If tokens, we need to switch it to samples. If samples, we need to rename it
There was a problem hiding this comment.
Updated in the latest commit.
|
@ShriyaPalsamudram @mmarcinkiewicz the latest run fails the RCP checker in the following test:
|
Further updating on this comment - I have ran the RCP checker against the latest compliance checker from PR 414 and it passed without any errors / warnings, so I'd say both PRs are ready to merge. |
3 participants
No description provided.